Introduction to Renjin The R interpreter in the JVM. Maarten-Jan Kallen

Transcription

1 Introduction to Renjin The R interpreter in the JVM Maarten-Jan Kallen mj@bedatadriven.com 1

2 Presentation outline What is Renjin? Renjin's design goals Other R engines Example Java application Example Google Appengine deployment Where to find information and help 2

3 3

4 4

5 5

6 To understand what Renjin is, you need to understand what (GNU) R is. 6

7 What is R? A dynamically interpreted, functional programming language specifically designed with statistical applications in mind R implements the grammar of the S language, with a few extensions (Chambers, 2008) There is no language reference, only the reference implementation referred to as GNU R The core of GNU R (i.e. the set of primitive functions) is written in C 7

8 How does R work? You type an expression, the language interpreter (or 'evaluator') parses and evaluates that expression, the interpreter prints out the result of the expression. = Read-Eval-Print-Loop ('REPL') Obviously, the 'evaluator' is the core element! 8

9 What makes up R? In his book Software for data analysis (2008) John Chambers states three concepts that are important to R: the language, the evaluator, and the management of memory for objects. For GNU R as a whole, I would add library management to this list 9

10 What is Renjin? Interpreter for the R language written in Java: primitive functions are written in Java JVM takes care of memory management (such as garbage collection) third-party Java libraries for things such as LAPACK/BLASS and file system access Like GNU R, it's open source 10

11 Overview of R engines (1) GNU R = standard interpreter written in C that is the reference implementation There are forks which use most of the same C code with some changes. Other engines are completely rewritten There is variation in commercial/academic as well as open/closed source Revolution R Enterprise and Oracle R Distribution are not much more than repackaged versions of GNU R 11

12 Overview of R engines (2) Engine What Who Language Source GNU R reference implementation R Core development team C open pqr fork of GNU R Radford Neal C open TERR closed-source interpreter that can be commercially licensed TIBCO C++ closed Riposte JIT and R interpreter Justin Talbot C++ open FastR Renjin academic research into formalizing the R language (TestR testing framework) and creating an interpreter commercially backed interpreter Purdue University Java open BeDataDriven Java open 12

13 Renjin's design goals Run R scripts and programs in the JVM: enable cloud computing to use scalability and reliability of Platform-as-a-Service providers like Google Appengine Improve performance and handle large(r) data sets Greater flexibility through abstraction of things such as data types and file system access High level of compatibility with GNU R and CRAN packages 13

14 We keep the best of R, namely the language itself, and improve the plumbing......which is pretty much everything else. 14

15 Free benefits Seamless integration with the rest of the enterprise: use tools in the Java ecosystem for profiling/debugging, project/dependency management, release/repository management, continuous integration, component lifecycle management, etc. More productive development cycle: no need to port a prototype written in R to Java or C++ for production use 15

16 No abstraction in GNU R GNU R version 3.1 introduces the anyna(x) function that does any(is.na(x)) anyna(x) is faster and uses less memory but, like paste0(), it introduces yet another function a discussion about the name of the function happened after it was implemented: (GNU R bug 15239) 16

17 Abstraction in Renjin (1) If you call is.na(x) on an atomic vector x of length > 100, it will return a view and then any() will operate on that view No copy of x is made No extra function is required All of the following statements also benefit from this abstraction: any(is.na(x) is.na(y)) all(!is.na(x)) x<-is.na(x);any(x) 17

18 How does it work in Renjin? Vector and AtomicVector are Java interfaces All vector types have abstract classes the view returned by is.na(x) is just a special implementation of the LogicalVector class called IsNaVector IsNaVector implements the Vector interface with an iselementna() method that can be used in a loop over the vector 18

19 Using abstraction, the performance issue of any(is.na(x)) is solved in the interpreter itself. 19

20 Abstraction in Renjin (2) Another example is x <- seq(from=1, to=1e6, by=1) No need to store this vector of length 1 million in memory so Renjin defines a DoubleSequence (or IntSequence) vector that stores it much like the command above. y <- x^2 still only returns a view: nothing gets copied. 20

21 Example 1: Renjin in a Java application Step 1: open up your favorite IDE (e.g. Eclipse or IntelliJ IDEA) Step 2: drop the Renjin Script Engine with dependencies JAR into your project library (download from Step 3: use the Java scripting API (i.e. import javax.script.*;) to evaluate R code within your Java program 21

22 Example 2: Renjin in Google Appengine git clone cd appengine-simple mvn appengine:devserver (this will download all necessary dependencies: mvn requires Maven 3.0+ to be installed) point your browser at and rejoice! 22

23 Will my R code work in Renjin? Probably not. 23

24 Current compatibility No graphics (not planned either) No S4 No MASS package No Rcpp (not planned either) About 73% of all primitive functions implemented Users cannot install GNU R packages, but we provide a repository of CRAN packages built for Renjin at 24

25 Using CRAN packages Just try library( pkgname ), no installation required, just an internet connection. 25

26 Conclusion Renjin is not yet for everyone, nor for exploratory analysis Best option for tight integration of R programs into a Java application No-brainer for deployment onto Google Appengine, Microsoft Azure, Red Hat OpenShift, etc. When we reach ~100% compatibility with GNU R, better performance will make Renjin a serious alternative We are looking for contributors and users (i.e. testers) 26

27 Where to find information Main website at for downloads, documentation and blog GitHub at for the source code and the issue tracker Documentation for Java developers at 27

28 Where to get help Google group (i.e. public mailing list) for developers (and users) at StackOverflow now has the [renjin] tag: Commercial support by BeDataDriven 28

29 How to stay up to date on Twitter Sign up for the newsletter or subscribe to the blog's RSS feed at 29